Schematic diagram of the current speech-segregation system. DNN = deep neural network, IBM = ideal binary mask.
Segregation of a HINT utterance from speech-shaped noise at −5 dB SNR. (a) Cochleagram of the utterance in quiet. (b) Cochleagram of the speech-plus-noise. (c) IBM. (d) Estimated IBM. (e) Cochleagram of the utterance following noise removal using the estimated IBM.
Pure-tone air-conduction audiometric thresholds for the listeners with sensorineural hearing impairment. Thresholds in right ears are represented by circles and those in left ears are represented by ×'s. Also displayed are listener ages in years and genders, as well as PTAs (in dB HL based on thresholds at 0.5, 1, and 2 kHz and pooled across ears). Listeners are numbered and arranged according to increasing PTA.
Mean HINT sentence component-word recognition scores for each listener. Normal-hearing listeners are represented by open symbols and hearing-impaired listeners are represented by filled symbols. Unprocessed conditions are represented by circles, and algorithm-processed conditions are represented by triangles. The upper panels display recognition in speech-shaped noise at three SNRs indicated, and the lower panels display recognition in multi-talker babble at three SNRs. The hearing-impaired listeners are numbered and plotted in order of increasing pure-tone average.
Group mean component-word recognition scores and standard errors for HINT sentences presented in speech-shaped noise (upper panels) and multi-talker babble (lower panels), at the SNRs indicated, for normal-hearing and hearing-impaired listeners, both prior to and following algorithm processing.
Article metrics loading...
Full text loading...